Conventional MRI-Based Semantic Features for Differentiation of Pediatric Medulloblastoma and Ependymoma in the Fourth Ventricle: Insights From a Multi-Center Retrospective Study.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
OBJECTIVE: Differentiating pediatric medulloblastoma (MB) from ependymoma (EA) in the fourth ventricle remains challenging due to overlapping clinical and imaging features. This study aimed to identify distinctive semantic features and develop a feature-based model for differentiating MB from EA using conventional MRI. METHODS: This multi-center retrospective MRI study enrolled 295 pediatric patients, including 184 MB and 111 EA cases, allocated to the training, internal validation set, and external testing set. Subsequently, 13 semantic features were extracted. After feature selection, quantitative parametric models and mixed parametric models were constructed and evaluated. Finally, three junior and three senior radiologists completed independent and model-assisted assessments. RESULTS: MB showed significantly greater left-right/upper-lower (0.98 vs. 0.76) and anterior-posterior/upper-lower (0.88 vs. 0.66) diameter ratios compared to EA (both p < 0.001). The pathognomonic "sea anemone sign" (100% specific for MB) occurred in 38.60% of MB cases. The support vector machine (SVM) model achieved optimal performance, with areas under the receiver operating characteristic curves (AUCs) of 0.946/0.921/0.915 and accuracies of 0.906/0.907/0.832 across training/internal validation/external testing sets. With SVM model assistance, diagnostic performance improved for both senior and junior radiologists across all datasets. In the external testing set, the AUCs increased to 0.849-0.893 for junior radiologists and 0.893-0.913 for senior radiologists, with the accuracies of 0.832-0.858 and 0.850-0.867, respectively. CONCLUSION: The sea anemone sign is highly specific for MB. The SVM model using conventional MRI semantic features achieved robust discrimination between MB and EA, significantly augmenting the diagnostic accuracy of radiologist assessment.